Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
Commits on main
tests: the loader ABI has 7 optional symbols now; dashboard phase sums may round to zero
af1cbb9 JustVugg committed 2 months ago
loader: coli_cuda_matmul_mxfp4 through the Windows DLL, so kimi_k3.exe links with CUDA_DLL=1
a73e470 JustVugg committed 2 months ago
Merge pull request #1406 from JustVugg/docs/windows-cuda-dll
adf009b Vincenzo Fornaro committed 2 months ago
ci: kimi_k3.exe with the DLL loader too, it links the same backend
0e3272a JustVugg committed 2 months ago
ci: qwen36.exe, the bare name has no alias and hits make's builtin rule
9ee0c02 JustVugg committed 2 months ago
ci: build the portable Windows CUDA DLL and publish it as an artifact (#1405)
1642b91 JustVugg committed 2 months ago
docs(windows): the CUDA DLL builds from a git clone in c/, not from the release zip (#1405)
fb365f6 JustVugg committed 2 months ago
fix(qwen36 tier): drive the production offer path in the test (#1391)
13157f0 Petteri Kosonen committed 2 months ago
fix(qwen36 tier): offer int8 experts on the decode path (#1391)
aa4629e Petteri Kosonen committed 2 months ago
feat(trunk): add TRUNK_RESIDENT_LAYERS=N mmap streaming for the dense trunk (#826)
d4788a7 SebaWag committed 2 months ago
feat(web): drop redundant Medium from GLM 5.3 reasoning selector
013342c Diego Moraes committed 2 months ago
Merge pull request #1394 from JustVugg/fix/cuda-prefix-per-row
1ccee43 Vincenzo Fornaro committed 2 months ago
Merge pull request #1393 from JustVugg/fix/glm53-thinking-split
cdc9975 Vincenzo Fornaro committed 2 months ago
Merge pull request #1244 from Unknown-Findout/fix/687-expert-sizing
c825e28 Vincenzo Fornaro committed 2 months ago
Merge pull request #1382 from dmoraesrs/feat/web-reasoning-levels
19d5adb Vincenzo Fornaro committed 2 months ago
cuda tier: price the VRAM prefix per row, not at the container's widest (#1351)
b44a034 JustVugg committed 2 months ago
Merge pull request #1392 from JustVugg/chore/issue-followups
9a87749 Vincenzo Fornaro committed 2 months ago
gateway: GLM-5.3 replies always start inside the think block (#1278)
2ef8dc5 JustVugg committed 2 months ago
Merge pull request #1387 from JustVugg/feat/kimi-qwen38-brain
aeee21e Vincenzo Fornaro committed 2 months ago
Merge pull request #1386 from JustVugg/feat/inkling-olmoe-hits
b5649e6 Vincenzo Fornaro committed 2 months ago
Follow-ups from the issue triage: probe visible in CI, doc cross-reference, three Azure rows
b729cc1 JustVugg committed 2 months ago
kimi_k3, qwen38: EMAP and HITS, so the Brain tab exists on both
7a8eec2 JustVugg committed 2 months ago
fix(qwen36 tier): qt_fill_wait returns only when the last upload is resident
df44b87 Courtney Andrew Richardson committed 2 months ago
qwen36 tier: drop the unused counter in the placement pricing loop
35fa507 Courtney Andrew Richardson committed 2 months ago
qwen36: free the warmstart int8 copy on ownership, not on format
84e99ab Courtney Andrew Richardson committed 2 months ago